Detecting OOD synthetic H&E images

This repository was made for the MoNuSeg diffusion model. This part of the code performs OOD filtering steps.

This repository is an extension of PatchCore as proposed in Roth et al. (2021), https://arxiv.org/abs/2106.08265 applied for OOD detection of synthetic histopathological images.

Step-by-step instruction

Data preparation

Download the data following the instructions in the MoNuSeg Dataset repository.

Then create a new data directory in this project with the name monuseg with train and test subdirectories. The train directory should only have real images in a folder named gt. The test directory can contain real images in a folder gt and synthetic images in a folder syn. The images should be of .png format and can either be copied or soft linked.

monuseg
├── train
│   └── gt
│       ├──img1.png
│       ├──img2.png
│       ├──
│       ├──dir1
│
├── test
    ├── gt
    │   ├──img1.png
    │   ├──img2.png
    │   ├──
    │   ├──dir2
    │
    └── syn
        ├──dir3
        ├──dir4
        ├──dir5

Installation

Build the docker image and then run it as follows.

docker build -t patchcore -f Dockerfile .

docker run --gpus all --rm -it -v $PATH_TO_DATASET:/mnt/dataset --name patchcore patchcore bash

Alternatively install the necessary packages manually without docker by using the requirements file.

Training

To train the patchcore model with H-Optimus-0 backbone use the run_patchcore script

python bin/run_patchcore.py 
results \
--gpu 0 --seed 0 \
--save_patchcore_model \
--log_group IM224_OPT_P01_D1024-1024_PS-3_AN-1_S0 --log_project MoNuSeg_Results \
patch_core -b optimus -le out --faiss_on_gpu \
--pretrain_embed_dimension 1024  --target_embed_dimension 1024 --anomaly_scorer_num_nn 1 --patchsize 3 \
sampler -p 0.1 approx_greedy_coreset \
dataset --resize 224 --imagesize 224 monuseg $PATH_TO_DATASET

$PATH_TO_DATASET refers to the folder that was created in the data preparation step with train adn test sets.

Set the --save_segmentation_images flag to save sample output anomaly maps.

Refer sample_training.sh for other examples with different backbones.

Evaluation

To evaluate on a pretrained model use the load_and_evaluate_patchcore script

python bin/load_and_evaluate_patchcore.py \
$RESULTS_PATH --gpu 0 --seed 0 --save_anomaly_scores \
patch_core_loader -p $MODEL_PATH --faiss_on_gpu \
dataset --resize 224 --imagesize 224 monuseg $PATH_TO_TEST_SET

Set the --save_segmentation_images flag to save sample output anomaly maps.

Hyperparameter Tuning

To Hyperparameter Tuning script is based on train and evaluation scripts. The model is trained on the training set once and evaluated on the test set multiple times. Use the on a pretrained model use the hyperparam_opt_patchcore script

Visualization

The anomaly scores for each set can be visualized using the analysis.ipynb script. It also displays sample images with corresponding anomaly values.

OOD scores	Samples

AUROC for hyperparameter tuning can be visualized using hyperparam.ipynb.